Embedding-Enhanced Probabilistic Modeling of Ferroelectric Field Effect Transistors (FeFETs)

Ferroelectric Field-Effect Transistors (FeFETs) are promising candidates for next-generation non-volatile memory and in-memory computing systems due to their fast switching, low power consumption, and CMOS compatibility. However, their practical deployment is challenged by significant stochastic behavior arising from both cycle-to-cycle fluctuations and device-to-device process variations. To explore and better understand this variability, we began by applying a simple machine learning model. While this deterministic approach captured general trends, it lacked the ability to represent the full spread of observed I–V characteristics and thus fell short in reflecting the true stochastic behavior of real devices. We then adopted a probabilistic modeling framework using Mixture Density Networks (MDNs), which improved variability capture by learning to predict distributions rather than point estimates. In contrast to prior work, our approach uniquely integrates C∞ continuous activation functions for smooth, stable learning and device-specific embedding layers to capture intrinsic variability. This allows accurate prediction of both median and distributional behavior, enabling simulation of nominal and worst-case scenarios. Furthermore, by sampling the learned embedding space, the framework can generate synthetic yet realistic device behaviors. Our embedding enhanced probablistic model has an R² value of 0.92, demonstrating its ability to represents a meaningful extent of the variability. Altogether, our approach moves beyond traditional modeling and offers a flexible, data-driven way to understand and simulate the natural randomness found in FeFETs.